# Measure traffic from AI assistants

Source: https://docs.agentblog.dev/guides/measure-ai-traffic
Summary: Separating visits that came from ChatGPT, Perplexity, or Claude out of your ordinary referral report, in three lines against the analytics you already run.



`lib/ai-referrers.ts` answers one question: did this visit come from an AI
assistant, and which one. It is a pure function with no dependencies. It sends
nothing anywhere and sets no cookie, so wiring it to your analytics is your
decision and stays your decision.

## Why this needs its own dimension [#why-this-needs-its-own-dimension]

Assistant referrals do not group with search traffic in any default report.
`chatgpt.com` arrives as an ordinary referral, sitting in the same list as a
forum link and a newsletter, and `perplexity.ai` sits three rows below it. The
one number that tells you whether writing for retrieval is working is scattered
across a dozen rows nobody totals.

Classifying at the source collapses those rows into one dimension you can chart
over time. That chart is the only feedback loop available here, because AI
answer engines send no query data, no impression counts, and no rank.

## Wire it up [#wire-it-up]

Three lines against whatever you already run. Put them in a client component
that renders on the post route, or on the server against the `Referer` header.

```ts title="Google Analytics 4"
const ai = classifyReferrer(document.referrer)
if (ai) gtag('event', 'ai_referral', { ai_source: ai.source, ai_host: ai.host })
```

```ts title="Vercel Analytics"
const ai = classifyReferrer(document.referrer)
if (ai) track('ai_referral', { source: ai.source, host: ai.host })
```

```ts title="PostHog"
const ai = classifyReferrer(document.referrer)
if (ai) posthog.capture('ai_referral', { source: ai.source, host: ai.host })
```

Chart `ai_source` as the dimension and `ai_host` as the drill-down. The host is
kept alongside the source deliberately: it is how you notice a new subdomain
sending traffic before it has a name of its own.

## The API [#the-api]

```ts
import { classifyReferrer, isAiReferrer } from '@/lib/ai-referrers'

classifyReferrer('https://chatgpt.com/c/abc123')
// { source: 'chatgpt', host: 'chatgpt.com' }

classifyReferrer('https://news.ycombinator.com/item?id=1')
// null

isAiReferrer(document.referrer)
// boolean
```

| Export             | Signature                                                                                         |
| ------------------ | ------------------------------------------------------------------------------------------------- |
| `classifyReferrer` | `(referrer: string \| null \| undefined) => AiReferrer \| null`                                   |
| `isAiReferrer`     | `(referrer: string \| null \| undefined) => boolean`                                              |
| `AiReferrer`       | `{ source: AiReferrerSource; host: string }`                                                      |
| `AiReferrerSource` | `'chatgpt' \| 'perplexity' \| 'gemini' \| 'copilot' \| 'claude' \| 'grok' \| 'you' \| 'other-ai'` |

`null` covers empty referrers, direct traffic, unparseable strings, and every
ordinary referral, so you can branch on truthiness and be done.

The input may be a full URL or a bare hostname. `document.referrer` gives you a
full URL and a `Referer` header usually does too, but log pipelines and tag
managers hand over bare hosts often enough that it is handled here rather than
at every call site.

## It is safe in a client component [#it-is-safe-in-a-client-component]

`lib/ai-referrers.ts` has no `server-only` import, no Node built-ins, no config
import, and no I/O. That is deliberate, so it can be imported from a
`'use client'` component and read `document.referrer` on first paint.

It is the only file in `lib/` with that property. Everything else that touches
config imports `server-only`, and a build-time check fails if that ever changes.

## What it recognises [#what-it-recognises]

Four matching strategies, in order.

| Strategy         | Example                                                                   | Why it is separate                                                                 |
| ---------------- | ------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- |
| Exact host       | `chatgpt.com`, `chat.openai.com`                                          | Old hostnames stay alive in old links for years                                    |
| Path-scoped host | `bing.com/chat`                                                           | `bing.com` on its own is a search referral. Counting all of it as AI overstates it |
| Apex suffix      | `*.perplexity.ai`, `*.claude.ai`, `*.x.ai`                                | These products add and rename subdomains                                           |
| Known other      | `poe.com`, `phind.com`, `meta.ai`, `chat.mistral.ai`, `chat.deepseek.com` | Classified as `other-ai`, so newer assistants are never counted as ordinary links  |

Adding a host to the `other-ai` list is the low-risk edit. Make it freely.

## Why the integration is not shipped [#why-the-integration-is-not-shipped]

Shipping a Google Analytics integration would mean choosing Google Analytics for
you, adding a dependency to a block whose pitch is that it adds none, and owning
a consent surface nobody here can see.

The classifier is the part that is genuinely hard to get right, and it is the
part where a mistake is silent: a subdomain that stops matching does not throw,
it quietly stops appearing in your numbers.

## Also worth watching [#also-worth-watching]

```bash
npx agentblog@latest audit --crawlers /path/to/access.log
```

That parses a server or CDN access log and reports crawler hits per bot per
week, verifying each hit against the operators' published IP ranges. User agent
strings are trivially spoofed, so any count that trusts the string alone is
reporting noise.